Deep learning-based identification of mixed-mode acoustic emission signals and damage coupling analysis in composite adhesive joints

Abstract

Acoustic emission (AE) enables real-time structural health monitoring with high sensitivity. However, overlapping signals from multiple concurrent sources—known as mixed-mode AE—pose major challenges for accurate damage classification. This paper presents a novel identification approach utilizing a deep learning-based ensemble method combined with tailored pre- and post-processing techniques. By segmenting time–frequency spectrograms of AE hits into frequency bands, the convolutional neural networks ensemble effectively extracts features to distinguish constituent damage modes within mixed signals. Among several architectures evaluated, DenseNet achieved the highest classification accuracy, exceeding 98.86% on independent test data. Information entropy analysis further confirmed clear spectral distinctions between pure-mode and mixed-mode AE signals, consistent with theoretical predictions. Model interpretability analysis elucidated the basis for model decisions and directions for improvement. Leveraging the reliable predictions, the coupling relationships between damage modes were deduced, and the finite element method was introduced to further explain the physical essence of coupling transformation, revealing the decisive role of the out-of-plane peeling stress in adhesive debonding and fiber breakage. Additionally, Gaussian process regression (GPR) models verified the existence of the mixed-mode signal formation pattern. Overall, the proposed method offers a robust solution for analyzing complex AE signals and provides new insights into the intrinsic mechanisms of AE activity in composite structures.